Heading Homing Simulation System Based on Image Intelligent Recognition
TELKOMNIKA, Vol.14, No.2A, June 2016, pp. 307~313
ISSN: 1693-6930, accredited A by DIKTI, Decree No: 58/DIKTI/Kep/2013
DOI: 10.12928/TELKOMNIKA.v14i2A.4346
307
Heading Homing Simulation System Based on Image
Intelligent Recognition
1
1
Zhao Ke* , Ziba Eslami
2
Research Center of Zhenjiang Watercraft College, Zhenjiang, China
2
Lockheed Martin Space Systems Company, Paraná, Brazil
*Corresponding author, e-mail: zkee@163.com
Abstract
Channel line recognition is one of extremely important technologies in intelligent driving field. In
recent years, machine vision has been the mainstream method to solve channel line recognition problems.
To overcome such deficiencies of existing channel line recognition algorithms as being complicated, slow
and short of robustness, the Thesis provides a new and rapid channel line recognition algorithm which
firstly obtains outline pixel of channel line through analysis on the images’ grayscale differences and then
applies B-Spline curve to fit the channel line profile, thus getting the final recognition effect picture. The
experiments show that excellent performances in both speed and recognition rate can result from the
algorithm. Besides, in embedded platform, the speed of the algorithm in the Thesis results in 12 frames per
second, which conforms to the real demands of intelligent driving.
Keywords: Machine vision; Channel line recognition; B-Spline; Curve fitting; Random sampling
consistency; Embedded system
1. Introduction
Ship Intelligent Handling and Control Simulation Platform is a simulation test platform
for automatic ship collision avoidance researches and shipping intellectualization and its control
module is one of the major innovative points of the Platform [1-3]. The main content of shipping
intellectualization includes realizing intelligent decision-making support and intelligent target
ship function for automatic ship collision avoidance. Besides, there are two automatic control
models for the ship, course and track control modules, which can achieve automatic ship
collision avoidance and automatic track monitoring respectively and be applied in algorithm
tests of both automatic ship collision avoidance and intelligent control [4-6]. As the development
of researches on shipping sciences and technologies has been focusing on improving shipping
safety and to realize automatic ship collision avoidance, minimize and completely eradicate
collision accidents by human factors from happening earlier, there is an increasing number of
shipping simulator projects such as automatic yacht handling and control training and simulation
platform entrusted by enterprises during the construction process of implementing platform
projects. For long-term considerations, the realization of shipping intellectualization requires that
the navigation equipment is equipped with a function of intelligent track control; and for shortterm considerations, the shipping intellectualization module of Ship Intelligent Handling and
Control Simulation Platform requires the realization of automatic ship control (intelligent control
in course and track) function [7-9].
2. Methods Applied in the Thesis
Aiming at the real demands of intelligent driving system, the Thesis provides a new
channel line recognition algorithm which has given considerations to speed, success rate of
tests and robustness. We firstly obtain the channel line profile information based on analysis on
the grayscale differences of images and then fit the channel line using B-spline curve to receive
the recognition result. The algorithm framework is as shown in Figure 1 [10-13].
Received February 3, 2016; Revised April 26, 2016; Accepted May 14, 2016
308
ISSN: 1693-6930
Read video images
Extract information of channel line profile
Fit the profile using B-Spline
Final recognition results
Figure 1. Algorithm flow diagram
2.1. Extraction of Channel Line Profile
Generally, channel lines are white or yellow lines which are outstanding under the
background of river surface and other ships in the images. In the Thesis, the profile is extracted
by adopting the characteristics of grayscale differences of channel lines. To keep a balance
among Channel R, G and B, in the Thesis, the value of grayscale is defined as:
(1)
In the Thesis, image on the same line is divided into several regions and each region
contains certain number of adjacent pixels. After obtaining the total grayscale of each pixel of
each region, we can analyze the grayscale differences of channel line in the images. The
and
in the equation refer to Region K and grayscale
definitions are as follows:
difference between Regions i+1 and i respectively. The value of region is determined by the
width of the channel line while in the Thesis it’s 6.
(2)
(3)
Figure 2. Original navigation channel and its grayscale difference analysis diagram
TELKOMNIKA Vol. 14, No. 2A, June 2016 : 307 – 313
TELKOMNIKA
ISSN: 1693-6930
309
Figure 2 has shown the read original grayscale image and the changing curve of region
i with D of certain line, the pixel of the line is marked with arrows in the original image. It can be
clearly seen from the figure that there is a pair of positive and negative values of D curve where
the channel lines lie.
The objective of the Thesis is to find out the channel line profile through these peak
values. In the algorithm, for the convenience of extracting peak values, several adjacent regions
are combined into one window where local maximum grayscale differences are to be found. The
area of the window is related to the width of the picture. In the Thesis, we set that there are 20
regions in each window, i.e. 120 pixels. Figure 3 is a schematic diagram of window, where I, i1…represent the regions in the window. The initial position of each window is located at the
right end of each line’s pixel, which moves to the left after algorithm and turns into Window’ until
the left end [13-16]. The algorithm procedures are as follows:
(1) Calculate all D’s values in the Window using Equation (3).
(2) Find out the maximal and minimal D values and record their positions; the profile
location of channel lines can be presented by a pair of adjacent positive and negative value.
(3) Filter the peak value. The positive peak value shall be larger than the minimum
grayscale difference between the channel line and background road. Through statistic
comparison, the threshold value in the Thesis is defined as 25.
(4) Move the window towards left and repeat Step 1 till the window lies at the left end of
the image.
Figure 3. Schematic diagram of window and its internal region’s structure
Figure 4 shows the recognition result of channel line profile. The original picture is the
first half in Figure 2.
Figure 4. Profile of the channel line
2.2. Matching of Channel Line Profile
After Step 1.1 above, the profile pixel of channel lines are already marked out, so
next step is to fit the profile of channel line using an appropriate mathematic model. In
Thesis, random sample consensus algorithm (RANSAC) based on B-Spline is adopted to fit
profile pixel of channel lines. B-Spline is a generalized form of Bessel function. Here,
the
the
the
we
Heading Homing Simulation System Based on Image Intelligent Recognition (Zhao Ke)
310
ISSN: 1693-6930
adopted the B-Spline curve to fit the profile for three times and obtain a mathematical
expression as follows:
(4)
Where
, , ,
and
are 4 control points of the curve.
The algorithm procedures are as follows:
(1) Random sampling. When randomly sample among profile pixel of channel lines, the
probability for a pixel to be selected is proportional to its grayscale difference D. Such a
sampling probability distribution should benefit to select the profile pixel of channel lines more
effectively and eliminate noise disturbance.
(2) Curve fitting. Conduct B-Spline fitting using least square method based on the
randomly sampled points.
(3) Fitting curve evaluation. Generally, the fitting results are evaluated by focusing on
obtaining the distances from all points to the fitted curve based on a fitting method of random
sample consensus algorithm. Given the huge calculations involved in this method, we here
adopt another simpler evaluation method. In most instances, channel lines in images are
relatively long curves with pretty small curvature. Through this characteristic, we can determine
a parallelogram area using fitting curve, as is shown in the dashed box of Figure 5. Then we
record the profile points S of channel lines within the area. The larger S is, the better the fitting
curve meet the demands. Besides, the width of the parallelogram is half of the window’s
aforesaid.
Figure 5. Schematic diagram for evaluating the quality of the fitted curve
The curve with maximum S, i.e. the objective curve, should be resulted from repeating
the above procedure of random sampling-fitting-algorithm till certain numbers of circulations are
finished. Figure 6 shows the effect picture after RANSAC fitting (the original image is the first
half in Figure 2). It can be found that, compared with previous profile extract effect picture
(Figure 4), RANSAC –based curve fitting can effectively extract channel line pixels conforming
to model features from images containing large amount of noises. Further channel line fitting
results will be discussed hereunder.
Figure 6. RANSAC fitting effect picture based on B-Spline model
TELKOMNIKA Vol. 14, No. 2A, June 2016 : 307 – 313
TELKOMNIKA
ISSN: 1693-6930
311
3. Experiment Result and Coresponding Analysis
In order to verify the rapid identification algorithm of channel lines provided in the
Thesis, we conduct several groups of experiments, the results of which prove not only the high
efficiency and accuracy of the algorithm but also its excellent effect on embedded equipment as
well as its application and real practice. The experiment contains two parts: the first part is
verification experiments which are many performed on desk computers to test the recognition
results and running speed of the algorithm; and the second part are transplant experiments
which applies the algorithm in embedded platform to test its efficiency in practical application.
3.1. Verification Experiments
To verify the effectiveness of the rapid channel line identification algorithm, a series of
images are randomly collected from the images shared by the Machine Vision Lab of The Ohio
State University, USA. The detailed numbers are as shown in Table 1 and the image resolution
is 640×480. Several channel line recognition effect picture are shown in Figure 7 which covers
various conditions including different road section, channel line type and backgrounds. In Figure
7, the left parts are the original images while the right parts are the recognized channel line
image. The figure has clearly proven that the algorithm in the Thesis can effectively recognize
channel line under conditions that there are barriers on the road or other sailing ships or the
channel lines are depicted using whether solid line or dotted line, straight line or curve, thus
presenting a good robustness.
Table 1. Source and number of the experiment images
Source of the sailling routes
wpt001
wpt002
wpt004
Number of image (frame)
1000
400
800
In the Thesis, the detailed algorithm accuracy rate is resulted from manually marking
the channel lines of the 15% images randomly selected from the test images. Several other
relatively typical algorithms are also selected to compare with the algorithm provided here,
including the channel line recognition method based on perspective transformation and Hough
transformation by Bertozzi and channel line recognition method [12] based on artificial intelligent
neutral network and curve fitting by Kim. The experiment results are as shown in Table 2. The
platform in the verification experiments is normal PC computer, the processor of which is Intel
Core 2 Dual-Core 1.8 GHz. To ensure fairness, all algorithms in the verification experiments are
performed only using C language and without further optimizations.
Table 2. Experiment and test results from different channel line algorithm
Name of algorithm
Bertozzi algorithm
Kim algorithm
Algorithm in the Thesis
Recognition accuracy rate
84.45%
86.71%
88.75%
Speed (frame per second)
25
24
29
It’s can be easily found out from contrast experiments that the channel line recognition
algorithm in the Thesis results in better effect compared with other two representative ones—the
former has higher accuracy and speed. The main causes include:
A more effective profile extraction method has been adopted in the algorithm. The
profile search methods based on regions are quicker and more successful in noise control.
Table 3 shows the contrast data between the profile search method in the Thesis and edge
inspection algorithm based on sobel operator [4], Roberts operator [5] and Prewitt operator [5].
In statistics, non-channel line pixels are all considered noise, thus the noise proportion in the
Thesis’s algorithm is apparently lower.
Heading Homing Simulation System Based on Image Intelligent Recognition (Zhao Ke)
312
ISSN: 1693-6930
Table 3. Noise proportions of various channel line profile extraction algorithm
Name of algorithm
Algorithm in the Thesis
Sobel operator
Perwitt operator
Roberts operator
Noise proportion
68.54%
80.72%
78.15%
76.63%
The quicker RANSAC fitting method is adopted in the Thesis. When sampling data with
RANSAC of the algorithm, the sampling probability is proportional to the grayscale difference D,
which makes it easier to extract the real channel line pixel, thus effectively reduce the number of
iterations—through the RANSAC method in the Thesis, only 35 iterations are required while 50
circulations are needed in the RANSAC adopted by Kim. As mentioned above, the calculation of
curve fitting is pretty large. Therefore, less iteration can greatly improve the running speed of the
algorithm.
3.2. Realization in Embedded Platform and the Test Results
After verifying the algorithm’s advantages in recognition accuracy and speed, we test
the algorithm in an embedded platform and optimize its performance. As both the internal ship
space and function are limited, embedded equipment has been the optimized choice for ship
intelligent driving system. Therefore, the running efficiency of channel line recognition algorithm
in embedded equipment shall decide its application future in real intelligent driving system. The
experiment platform in the Thesis is an experiment panel based on TI OMAP4430 chip, an
embedded processing chip equipped with dual-core ARM Cortex and have been widely applied
in various mobile equipment (such as ship-load navigation equipment, e-book reader and smart
phone etc.), and 512MB SDRAM. Figure 7 shows the internal structure [17] of OMAP 4430 chip.
Figure 7. Internal structure of OMAP 4430 chip
Given the processing feature of the embedded platform, the following optimizing
procedures are conducted in transplanting the algorithm:
(1) Use fixed-point number as the calculation form. The OMAP chip is more effective in
processing fixed-point number calculation than floating-point number calculation, so where there
require floating-point number calculation shall be transferred into fixed-point numbers.
(2) Make full use of the DSP calculation module in OMAP 4430 chip—use DSP to
speed up the curve fitting part with larger data calculations (other parts shall be processed by
ARM dual-core processor as their if-else is multiple—making full use of the hardware property of
the platform).
Table 4. Test results on embedded platform
Name of algorithm
Algorithm in the Thesis
Speed before optimizing
9 frames per second
TELKOMNIKA Vol. 14, No. 2A, June 2016 : 307 – 313
Speed after optimizing
12 frames per second
TELKOMNIKA
ISSN: 1693-6930
313
As shown in Table 4, after being optimized, the running speed of the algorithm in
embedded platform can be up to 12 frames per second, 33% quicker than the speed before
optimizing. Generally, the driving speed of ship on highways is around 100km/h (≈28m/s) and
lowers on municipal non-overhead roads. Under conditions that the visibility is not too bad
(except extreme atrocious weather such as heavy fog, rain and snow etc.), a frame of image
can totally grasp the road condition 30m ahead. Therefore, the rapid channel line recognition
provided in the Thesis can keep a relatively rapid running speed in embedded equipment and
can satisfy the demand of real intelligent driving system.
4. Conclusion
The Thesis provides a rapid channel line recognition algorithm on the basis of machine
vision which firstly outlines the profile pixel of channel line using the grayscale differences
between channel lines and background roads and then fit the profile pixel using RANSAC based
on B-Spline model. From the results of several verification experiments, we find that the
recognition rate and speed of this method is up to 88.75% and 29 frames per second
respectively, which has obvious performance advantages compared with the classic algorithm in
the domain. Besides, aiming at the demands of real ship intelligent driving system, the algorithm
is also tested and optimized on embedded platform. The optimized speed can reach 12 frames
per second, which has huge application potential in intelligent sailing domain.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
Soni RA. A Study on Android Application Development. Journal of Telematics and Informatics. 2013;
1(2): 89-96.
Lv Z, Li X, Zhang B. Managing Big City Information Based on WebVRGIS. IEEE Access. 2016; 4:
407-415.
Corbier C, Badaoui ME, Ugalde HMR. Huberian Approach for Reduced Order ARMA Modeling of
Neurodegenerative Disorder Signal. Signal Processing. 2015, 113: 273-284.
Su S, Wang S. Simple Monitoring Network System of Wireless Sensor Network. Bulletin of Electrical
Engineering and Informatics. 2012; 1(4): 251-254.
Jiang D, Xu Z, Lv Z. A multicast delivery approach with minimum energy consumption for wireless
multi-hop networks. Telecommunication Systems. 2015; 1-12.
Liu Y, Yang J, Meng Q. Stereoscopic image quality assessment method based on binocular
combination saliency model. Signal Processing. 2016; 125: 237-248.
Lv Z, Yin T, Han Y. WebVR-web virtual reality engine based on P2P network. Journal of Networks.
2011; 6(7): 990-998.
Yang J, Zhou J, Lv Z. A Real-Time Monitoring System of Industry Carbon Monoxide Based on
Wireless Sensor Networks. Sensors. 2015; 15(11): 29535-29546.
Jiang D, Ying X, Han Y. Collaborative multi-hop routing in cognitive wireless networks. Wireless
Personal Communications. 2016; 86(2): 901-923.
M Abdar, SRN Kalhori, T Sutikno, IMI Subroto, G Arji. Comparing Performance of Data Mining
Algorithms in Prediction Heart Diseases. International Journal of Electrical and Computer
Engineering (IJECE). 2015; 5(6): 1569-1576.
Lv Z, Halawani A, Feng S. Multimodal hand and foot gesture interaction for handheld devices. ACM
Transactions on Multimedia Computing, Communications, and Applications (TOMM). 2014; 11(1s):
120-128.
Cui Y, Zhang S, Chen Z, Zheng W. A New Digital Image Hiding Algorithm Based on Wavelet Packet
Transform and Singular Value Decomposition. TELKOMNIKA Indonesian Journal of Electrical
Engineering. 2014; 12(7): 5408-5413.
Pennestrì E, Belfiore NP. On Crossley's contribution to the development of graph based algorithms
for the analysis of mechanisms and gear trains. Mechanism & Machine Theory. 2015; 89: 92-106.
Lv Z, Halawani A, Fen S. Touch-less Interactive Augmented Reality Game on Vision Based
Wearable Device. Personal and Ubiquitous Computing. 2015; 19(3): 551-567.
Ontrup J, Wersing H, Ritter H. A computational feature binding model of human texture perception.
Cognitive Processing. 2004; 5(1): 31-44.
Degui Zeng, Yishuang Geng. Content distribution mechanism in mobile P2P network. Journal of
Networks. 2014; 9(5): 1229-1236.
Heading Homing Simulation System Based on Image Intelligent Recognition (Zhao Ke)
ISSN: 1693-6930, accredited A by DIKTI, Decree No: 58/DIKTI/Kep/2013
DOI: 10.12928/TELKOMNIKA.v14i2A.4346
307
Heading Homing Simulation System Based on Image
Intelligent Recognition
1
1
Zhao Ke* , Ziba Eslami
2
Research Center of Zhenjiang Watercraft College, Zhenjiang, China
2
Lockheed Martin Space Systems Company, Paraná, Brazil
*Corresponding author, e-mail: zkee@163.com
Abstract
Channel line recognition is one of extremely important technologies in intelligent driving field. In
recent years, machine vision has been the mainstream method to solve channel line recognition problems.
To overcome such deficiencies of existing channel line recognition algorithms as being complicated, slow
and short of robustness, the Thesis provides a new and rapid channel line recognition algorithm which
firstly obtains outline pixel of channel line through analysis on the images’ grayscale differences and then
applies B-Spline curve to fit the channel line profile, thus getting the final recognition effect picture. The
experiments show that excellent performances in both speed and recognition rate can result from the
algorithm. Besides, in embedded platform, the speed of the algorithm in the Thesis results in 12 frames per
second, which conforms to the real demands of intelligent driving.
Keywords: Machine vision; Channel line recognition; B-Spline; Curve fitting; Random sampling
consistency; Embedded system
1. Introduction
Ship Intelligent Handling and Control Simulation Platform is a simulation test platform
for automatic ship collision avoidance researches and shipping intellectualization and its control
module is one of the major innovative points of the Platform [1-3]. The main content of shipping
intellectualization includes realizing intelligent decision-making support and intelligent target
ship function for automatic ship collision avoidance. Besides, there are two automatic control
models for the ship, course and track control modules, which can achieve automatic ship
collision avoidance and automatic track monitoring respectively and be applied in algorithm
tests of both automatic ship collision avoidance and intelligent control [4-6]. As the development
of researches on shipping sciences and technologies has been focusing on improving shipping
safety and to realize automatic ship collision avoidance, minimize and completely eradicate
collision accidents by human factors from happening earlier, there is an increasing number of
shipping simulator projects such as automatic yacht handling and control training and simulation
platform entrusted by enterprises during the construction process of implementing platform
projects. For long-term considerations, the realization of shipping intellectualization requires that
the navigation equipment is equipped with a function of intelligent track control; and for shortterm considerations, the shipping intellectualization module of Ship Intelligent Handling and
Control Simulation Platform requires the realization of automatic ship control (intelligent control
in course and track) function [7-9].
2. Methods Applied in the Thesis
Aiming at the real demands of intelligent driving system, the Thesis provides a new
channel line recognition algorithm which has given considerations to speed, success rate of
tests and robustness. We firstly obtain the channel line profile information based on analysis on
the grayscale differences of images and then fit the channel line using B-spline curve to receive
the recognition result. The algorithm framework is as shown in Figure 1 [10-13].
Received February 3, 2016; Revised April 26, 2016; Accepted May 14, 2016
308
ISSN: 1693-6930
Read video images
Extract information of channel line profile
Fit the profile using B-Spline
Final recognition results
Figure 1. Algorithm flow diagram
2.1. Extraction of Channel Line Profile
Generally, channel lines are white or yellow lines which are outstanding under the
background of river surface and other ships in the images. In the Thesis, the profile is extracted
by adopting the characteristics of grayscale differences of channel lines. To keep a balance
among Channel R, G and B, in the Thesis, the value of grayscale is defined as:
(1)
In the Thesis, image on the same line is divided into several regions and each region
contains certain number of adjacent pixels. After obtaining the total grayscale of each pixel of
each region, we can analyze the grayscale differences of channel line in the images. The
and
in the equation refer to Region K and grayscale
definitions are as follows:
difference between Regions i+1 and i respectively. The value of region is determined by the
width of the channel line while in the Thesis it’s 6.
(2)
(3)
Figure 2. Original navigation channel and its grayscale difference analysis diagram
TELKOMNIKA Vol. 14, No. 2A, June 2016 : 307 – 313
TELKOMNIKA
ISSN: 1693-6930
309
Figure 2 has shown the read original grayscale image and the changing curve of region
i with D of certain line, the pixel of the line is marked with arrows in the original image. It can be
clearly seen from the figure that there is a pair of positive and negative values of D curve where
the channel lines lie.
The objective of the Thesis is to find out the channel line profile through these peak
values. In the algorithm, for the convenience of extracting peak values, several adjacent regions
are combined into one window where local maximum grayscale differences are to be found. The
area of the window is related to the width of the picture. In the Thesis, we set that there are 20
regions in each window, i.e. 120 pixels. Figure 3 is a schematic diagram of window, where I, i1…represent the regions in the window. The initial position of each window is located at the
right end of each line’s pixel, which moves to the left after algorithm and turns into Window’ until
the left end [13-16]. The algorithm procedures are as follows:
(1) Calculate all D’s values in the Window using Equation (3).
(2) Find out the maximal and minimal D values and record their positions; the profile
location of channel lines can be presented by a pair of adjacent positive and negative value.
(3) Filter the peak value. The positive peak value shall be larger than the minimum
grayscale difference between the channel line and background road. Through statistic
comparison, the threshold value in the Thesis is defined as 25.
(4) Move the window towards left and repeat Step 1 till the window lies at the left end of
the image.
Figure 3. Schematic diagram of window and its internal region’s structure
Figure 4 shows the recognition result of channel line profile. The original picture is the
first half in Figure 2.
Figure 4. Profile of the channel line
2.2. Matching of Channel Line Profile
After Step 1.1 above, the profile pixel of channel lines are already marked out, so
next step is to fit the profile of channel line using an appropriate mathematic model. In
Thesis, random sample consensus algorithm (RANSAC) based on B-Spline is adopted to fit
profile pixel of channel lines. B-Spline is a generalized form of Bessel function. Here,
the
the
the
we
Heading Homing Simulation System Based on Image Intelligent Recognition (Zhao Ke)
310
ISSN: 1693-6930
adopted the B-Spline curve to fit the profile for three times and obtain a mathematical
expression as follows:
(4)
Where
, , ,
and
are 4 control points of the curve.
The algorithm procedures are as follows:
(1) Random sampling. When randomly sample among profile pixel of channel lines, the
probability for a pixel to be selected is proportional to its grayscale difference D. Such a
sampling probability distribution should benefit to select the profile pixel of channel lines more
effectively and eliminate noise disturbance.
(2) Curve fitting. Conduct B-Spline fitting using least square method based on the
randomly sampled points.
(3) Fitting curve evaluation. Generally, the fitting results are evaluated by focusing on
obtaining the distances from all points to the fitted curve based on a fitting method of random
sample consensus algorithm. Given the huge calculations involved in this method, we here
adopt another simpler evaluation method. In most instances, channel lines in images are
relatively long curves with pretty small curvature. Through this characteristic, we can determine
a parallelogram area using fitting curve, as is shown in the dashed box of Figure 5. Then we
record the profile points S of channel lines within the area. The larger S is, the better the fitting
curve meet the demands. Besides, the width of the parallelogram is half of the window’s
aforesaid.
Figure 5. Schematic diagram for evaluating the quality of the fitted curve
The curve with maximum S, i.e. the objective curve, should be resulted from repeating
the above procedure of random sampling-fitting-algorithm till certain numbers of circulations are
finished. Figure 6 shows the effect picture after RANSAC fitting (the original image is the first
half in Figure 2). It can be found that, compared with previous profile extract effect picture
(Figure 4), RANSAC –based curve fitting can effectively extract channel line pixels conforming
to model features from images containing large amount of noises. Further channel line fitting
results will be discussed hereunder.
Figure 6. RANSAC fitting effect picture based on B-Spline model
TELKOMNIKA Vol. 14, No. 2A, June 2016 : 307 – 313
TELKOMNIKA
ISSN: 1693-6930
311
3. Experiment Result and Coresponding Analysis
In order to verify the rapid identification algorithm of channel lines provided in the
Thesis, we conduct several groups of experiments, the results of which prove not only the high
efficiency and accuracy of the algorithm but also its excellent effect on embedded equipment as
well as its application and real practice. The experiment contains two parts: the first part is
verification experiments which are many performed on desk computers to test the recognition
results and running speed of the algorithm; and the second part are transplant experiments
which applies the algorithm in embedded platform to test its efficiency in practical application.
3.1. Verification Experiments
To verify the effectiveness of the rapid channel line identification algorithm, a series of
images are randomly collected from the images shared by the Machine Vision Lab of The Ohio
State University, USA. The detailed numbers are as shown in Table 1 and the image resolution
is 640×480. Several channel line recognition effect picture are shown in Figure 7 which covers
various conditions including different road section, channel line type and backgrounds. In Figure
7, the left parts are the original images while the right parts are the recognized channel line
image. The figure has clearly proven that the algorithm in the Thesis can effectively recognize
channel line under conditions that there are barriers on the road or other sailing ships or the
channel lines are depicted using whether solid line or dotted line, straight line or curve, thus
presenting a good robustness.
Table 1. Source and number of the experiment images
Source of the sailling routes
wpt001
wpt002
wpt004
Number of image (frame)
1000
400
800
In the Thesis, the detailed algorithm accuracy rate is resulted from manually marking
the channel lines of the 15% images randomly selected from the test images. Several other
relatively typical algorithms are also selected to compare with the algorithm provided here,
including the channel line recognition method based on perspective transformation and Hough
transformation by Bertozzi and channel line recognition method [12] based on artificial intelligent
neutral network and curve fitting by Kim. The experiment results are as shown in Table 2. The
platform in the verification experiments is normal PC computer, the processor of which is Intel
Core 2 Dual-Core 1.8 GHz. To ensure fairness, all algorithms in the verification experiments are
performed only using C language and without further optimizations.
Table 2. Experiment and test results from different channel line algorithm
Name of algorithm
Bertozzi algorithm
Kim algorithm
Algorithm in the Thesis
Recognition accuracy rate
84.45%
86.71%
88.75%
Speed (frame per second)
25
24
29
It’s can be easily found out from contrast experiments that the channel line recognition
algorithm in the Thesis results in better effect compared with other two representative ones—the
former has higher accuracy and speed. The main causes include:
A more effective profile extraction method has been adopted in the algorithm. The
profile search methods based on regions are quicker and more successful in noise control.
Table 3 shows the contrast data between the profile search method in the Thesis and edge
inspection algorithm based on sobel operator [4], Roberts operator [5] and Prewitt operator [5].
In statistics, non-channel line pixels are all considered noise, thus the noise proportion in the
Thesis’s algorithm is apparently lower.
Heading Homing Simulation System Based on Image Intelligent Recognition (Zhao Ke)
312
ISSN: 1693-6930
Table 3. Noise proportions of various channel line profile extraction algorithm
Name of algorithm
Algorithm in the Thesis
Sobel operator
Perwitt operator
Roberts operator
Noise proportion
68.54%
80.72%
78.15%
76.63%
The quicker RANSAC fitting method is adopted in the Thesis. When sampling data with
RANSAC of the algorithm, the sampling probability is proportional to the grayscale difference D,
which makes it easier to extract the real channel line pixel, thus effectively reduce the number of
iterations—through the RANSAC method in the Thesis, only 35 iterations are required while 50
circulations are needed in the RANSAC adopted by Kim. As mentioned above, the calculation of
curve fitting is pretty large. Therefore, less iteration can greatly improve the running speed of the
algorithm.
3.2. Realization in Embedded Platform and the Test Results
After verifying the algorithm’s advantages in recognition accuracy and speed, we test
the algorithm in an embedded platform and optimize its performance. As both the internal ship
space and function are limited, embedded equipment has been the optimized choice for ship
intelligent driving system. Therefore, the running efficiency of channel line recognition algorithm
in embedded equipment shall decide its application future in real intelligent driving system. The
experiment platform in the Thesis is an experiment panel based on TI OMAP4430 chip, an
embedded processing chip equipped with dual-core ARM Cortex and have been widely applied
in various mobile equipment (such as ship-load navigation equipment, e-book reader and smart
phone etc.), and 512MB SDRAM. Figure 7 shows the internal structure [17] of OMAP 4430 chip.
Figure 7. Internal structure of OMAP 4430 chip
Given the processing feature of the embedded platform, the following optimizing
procedures are conducted in transplanting the algorithm:
(1) Use fixed-point number as the calculation form. The OMAP chip is more effective in
processing fixed-point number calculation than floating-point number calculation, so where there
require floating-point number calculation shall be transferred into fixed-point numbers.
(2) Make full use of the DSP calculation module in OMAP 4430 chip—use DSP to
speed up the curve fitting part with larger data calculations (other parts shall be processed by
ARM dual-core processor as their if-else is multiple—making full use of the hardware property of
the platform).
Table 4. Test results on embedded platform
Name of algorithm
Algorithm in the Thesis
Speed before optimizing
9 frames per second
TELKOMNIKA Vol. 14, No. 2A, June 2016 : 307 – 313
Speed after optimizing
12 frames per second
TELKOMNIKA
ISSN: 1693-6930
313
As shown in Table 4, after being optimized, the running speed of the algorithm in
embedded platform can be up to 12 frames per second, 33% quicker than the speed before
optimizing. Generally, the driving speed of ship on highways is around 100km/h (≈28m/s) and
lowers on municipal non-overhead roads. Under conditions that the visibility is not too bad
(except extreme atrocious weather such as heavy fog, rain and snow etc.), a frame of image
can totally grasp the road condition 30m ahead. Therefore, the rapid channel line recognition
provided in the Thesis can keep a relatively rapid running speed in embedded equipment and
can satisfy the demand of real intelligent driving system.
4. Conclusion
The Thesis provides a rapid channel line recognition algorithm on the basis of machine
vision which firstly outlines the profile pixel of channel line using the grayscale differences
between channel lines and background roads and then fit the profile pixel using RANSAC based
on B-Spline model. From the results of several verification experiments, we find that the
recognition rate and speed of this method is up to 88.75% and 29 frames per second
respectively, which has obvious performance advantages compared with the classic algorithm in
the domain. Besides, aiming at the demands of real ship intelligent driving system, the algorithm
is also tested and optimized on embedded platform. The optimized speed can reach 12 frames
per second, which has huge application potential in intelligent sailing domain.
References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
Soni RA. A Study on Android Application Development. Journal of Telematics and Informatics. 2013;
1(2): 89-96.
Lv Z, Li X, Zhang B. Managing Big City Information Based on WebVRGIS. IEEE Access. 2016; 4:
407-415.
Corbier C, Badaoui ME, Ugalde HMR. Huberian Approach for Reduced Order ARMA Modeling of
Neurodegenerative Disorder Signal. Signal Processing. 2015, 113: 273-284.
Su S, Wang S. Simple Monitoring Network System of Wireless Sensor Network. Bulletin of Electrical
Engineering and Informatics. 2012; 1(4): 251-254.
Jiang D, Xu Z, Lv Z. A multicast delivery approach with minimum energy consumption for wireless
multi-hop networks. Telecommunication Systems. 2015; 1-12.
Liu Y, Yang J, Meng Q. Stereoscopic image quality assessment method based on binocular
combination saliency model. Signal Processing. 2016; 125: 237-248.
Lv Z, Yin T, Han Y. WebVR-web virtual reality engine based on P2P network. Journal of Networks.
2011; 6(7): 990-998.
Yang J, Zhou J, Lv Z. A Real-Time Monitoring System of Industry Carbon Monoxide Based on
Wireless Sensor Networks. Sensors. 2015; 15(11): 29535-29546.
Jiang D, Ying X, Han Y. Collaborative multi-hop routing in cognitive wireless networks. Wireless
Personal Communications. 2016; 86(2): 901-923.
M Abdar, SRN Kalhori, T Sutikno, IMI Subroto, G Arji. Comparing Performance of Data Mining
Algorithms in Prediction Heart Diseases. International Journal of Electrical and Computer
Engineering (IJECE). 2015; 5(6): 1569-1576.
Lv Z, Halawani A, Feng S. Multimodal hand and foot gesture interaction for handheld devices. ACM
Transactions on Multimedia Computing, Communications, and Applications (TOMM). 2014; 11(1s):
120-128.
Cui Y, Zhang S, Chen Z, Zheng W. A New Digital Image Hiding Algorithm Based on Wavelet Packet
Transform and Singular Value Decomposition. TELKOMNIKA Indonesian Journal of Electrical
Engineering. 2014; 12(7): 5408-5413.
Pennestrì E, Belfiore NP. On Crossley's contribution to the development of graph based algorithms
for the analysis of mechanisms and gear trains. Mechanism & Machine Theory. 2015; 89: 92-106.
Lv Z, Halawani A, Fen S. Touch-less Interactive Augmented Reality Game on Vision Based
Wearable Device. Personal and Ubiquitous Computing. 2015; 19(3): 551-567.
Ontrup J, Wersing H, Ritter H. A computational feature binding model of human texture perception.
Cognitive Processing. 2004; 5(1): 31-44.
Degui Zeng, Yishuang Geng. Content distribution mechanism in mobile P2P network. Journal of
Networks. 2014; 9(5): 1229-1236.
Heading Homing Simulation System Based on Image Intelligent Recognition (Zhao Ke)